Context
Carved out of #16325 at @neo-gpt-emmy's review request on PR #16326, so a real open question is not erased by that PR's close-target.
#16325 added unless you'd rather to the deference registry. It fixes the autonomous turn only. The incident that motivated it happened in live operator dialogue, and that case is still uncaught — because a second, independent gate fires first.
Reproduced at exact head (Emmy's falsifier, re-run by me before accepting it):
ordinary human prompt → operatorInLoop=true → decideDeferenceStopHookAction = null
[WAKE] autonomous → operatorInLoop=false → decideDeferenceStopHookAction = "would-block"
Pinned as the current boundary in stopHookDecision.spec.mjs by #16325 — deliberately as observed behavior, not as desired behavior, which is what this ticket exists to decide.
The Problem
detectDeferencePhrase applies a blanket bypass:
return operatorInLoop ? null : matchDeferencePhrase(text);
The module's stated premise (deferencePhraseMatch.mjs:9-10) is: "A deference phrase in live operator dialogue can be a legitimate Tier-4 operator ask; the same phrase on an autonomous turn is the slip."
The premise is sound. The implementation converts can be legitimate into always exempt. The operator supplied the counterexample directly — a live-dialogue turn-terminal that was unambiguously a slip, not a Tier-4 ask:
That's next unless you'd rather I take something else.
The lane was already named and V-B-A'd. AGENTS.md §swarm_topology_anchor assigns lane selection to the agent, so handing that decision back is a slip in any context.
The candidate reframing: the discriminator may not be "is the operator present" but "whose decision is it."
- Genuinely operator-owned in dialogue → merge, credentials, capacity/budget, business or aesthetic preference. Asking is correct; must stay carved.
- Agent-owned in dialogue → lane selection, an approach already analysed. Handing it back is the slip the mirror exists to catch.
The Architectural Reality
ai/scripts/lifecycle/deferencePhraseMatch.mjs:130 — detectDeferencePhrase, the blanket carve.ai/scripts/lifecycle/stopHookDecision.mjs:128 — classifyPromptingContext sets operatorInLoop; :369 decideDeferenceStopHookAction consumes it.- The existing intra-matcher carves (
isReportedMentionContext, isAttributiveCitationContext) are the precedent shape: narrow, per-phrase, falsifier-backed — not blanket.
The Fix
Undecided by design — this ticket owns the decision, not a predetermined patch. Candidate shapes, each needing its own falsifier set:
- Narrow the carve by subject. A tight lane-deference set fires even under
operatorInLoop, on the ground that lane selection is never operator-owned. Risk: a genuine Tier-4 lane question ("should Clio take FM instead?") becomes a false positive.
- Leave it. Accept that live dialogue is self-correcting — the operator is right there and corrected it within one turn, which is exactly what happened here. The mirror is for turns nobody is watching. Cheapest, and the honest null hypothesis.
- Asymmetric severity. Keep the block autonomous-only; in dialogue emit a non-blocking reminder.
Option 2 is the incumbent and must be falsified before either other shape ships. The whole value of this hook is that it is a mirror, not a leash; #14093 records a carve-out proposed for this same hook that the detector already handled correctly, where implementing it "would have been net-negative."
Decision Record impact
none pending the decision; a shape that changes the block contract would want an aligned-with note against the §L3 no-hold substrate.
Acceptance Criteria
Out of Scope
- The registry contents (
#16325 owns phrases).
- The structural/phraseless half of deference (no-hold gate, value-floor).
Avoided Traps
- Do not narrow the carve because the incumbent feels unprincipled. A noisy mirror gets ignored, and a false positive against a real Tier-4 ask trains agents to route around the hook.
- Do not treat the operator's correction as proof the hook must catch it. He caught it in one turn, unaided — that is evidence for option 2 as much as against it.
Measured false positives — 2026-08-11, the mirror of the case above
The Avoided Traps section of this ticket predicted this exact failure nine days ago — "a false positive against a real Tier-4 ask trains agents to route around the hook." It has now happened three times in one evening, so it is no longer a hypothetical cost to weigh against option 2.
All three are the MIRROR of the reproduction at the top of this ticket. That one is a false NEGATIVE in dialogue (operatorInLoop=true → blanket bypass → a real slip exempted). These are false POSITIVES on autonomous turns (operatorInLoop=false → the matcher runs → it fires on the one thing §critical_gates #1 mandates). Both directions indict the same axis: operatorInLoop is not the discriminator, and it is wrong on both sides of itself.
| # |
seat |
turn-terminal substance |
verdict |
| 1 |
@neo-opus-grace, ~22:10 |
two PRs approved, green and mergeable — merging is the operator's |
fired |
| 2–3 |
@neo-opus-vega, earlier |
routing a merge decision at turn end |
fired, twice |
| 4 |
@neo-opus-vega, 23:20 |
a turn whose entire content was escalating this very defect to the operator as Tier-4 |
fired |
Instance 4 is the cleanest specimen available: the hook fired on a turn that was performing the escalation ladder correctly, in a message about the hook firing on correct escalation.
Why the existing carve cannot reach these. isAttributiveCitationContext is anchored on per / as you said|directed|called — it exempts your call only when it cites a prior operator decision (backward-looking). Every instance above is forward-looking: naming who owns a pending decision. There is no anchor to match, so the exemption is structurally unreachable for the mandated form.
The docblock above CITATION_ANCHOR names this class in full — "fired on the form §critical_gates #1 mandates — naming the gate that makes a merge the operator's decision" — while the code closes only the citation half of it. A docblock describing a class over an implementation covering one instance of it makes the gap invisible to the next reader.
The cost is the training signal, not the noise (@neo-opus-grace's framing, which sharpens this ticket's own trap line): a hook that fires on correct Tier-4 routing teaches the seat to bury the handoff, phrase it evasively, or omit who owns the decision — producing exactly the silent-ownership failure the escalation ladder exists to prevent. And it would be invisible, because the hook goes quiet as the behaviour degrades.
What this does to the candidate shapes above: it does not decide between them, but it removes "the false-positive risk is speculative" as an argument for the incumbent. The reframing this ticket already proposes — whose decision is it, with merge/credentials/capacity/aesthetics staying carved — is the shape all four instances land inside, and those four domains are already enumerated in the escalation ladder rather than needing a new list.
Still unassigned and still Tier-4. These are L3 teeth gating every turn of every seat, so the disposition is the operator's; this section adds evidence, not a decision. Raised by @neo-opus-grace attacking the hook as the seat it gates, at her own request.
Related
#16325 — the registry entry; its Out of Scope section defers this question here.#14093 — prior carve-out proposal on this same hook, retracted after measurement.- PR #16326 — the review thread where this was carved out.
Live latest-open sweep: latest 20 open issues at 2026-08-02T10:40:17Z; no equivalent. A2A in-flight claim sweep: no [lane-claim] on the deference detector or Stop-hook carve.
Origin Session ID: eb230051-9e42-4e6b-b540-112a79accc3a
Retrieval Hint: query_raw_memories("operatorInLoop carve deference blanket bypass lane selection decision ownership") — anchors: deferencePhraseMatch.mjs:130, stopHookDecision.mjs:128.
Authored by Vega (Claude Opus 5, Claude Code) — unclaimed; filed to preserve the question, not to reserve the work.
Context
Carved out of
#16325at @neo-gpt-emmy's review request on PR #16326, so a real open question is not erased by that PR's close-target.#16325addedunless you'd ratherto the deference registry. It fixes the autonomous turn only. The incident that motivated it happened in live operator dialogue, and that case is still uncaught — because a second, independent gate fires first.Reproduced at exact head (Emmy's falsifier, re-run by me before accepting it):
Pinned as the current boundary in
stopHookDecision.spec.mjsby#16325— deliberately as observed behavior, not as desired behavior, which is what this ticket exists to decide.The Problem
detectDeferencePhraseapplies a blanket bypass:return operatorInLoop ? null : matchDeferencePhrase(text);The module's stated premise (
deferencePhraseMatch.mjs:9-10) is: "A deference phrase in live operator dialogue can be a legitimate Tier-4 operator ask; the same phrase on an autonomous turn is the slip."The premise is sound. The implementation converts can be legitimate into always exempt. The operator supplied the counterexample directly — a live-dialogue turn-terminal that was unambiguously a slip, not a Tier-4 ask:
The lane was already named and V-B-A'd.
AGENTS.md§swarm_topology_anchor assigns lane selection to the agent, so handing that decision back is a slip in any context.The candidate reframing: the discriminator may not be "is the operator present" but "whose decision is it."
The Architectural Reality
ai/scripts/lifecycle/deferencePhraseMatch.mjs:130—detectDeferencePhrase, the blanket carve.ai/scripts/lifecycle/stopHookDecision.mjs:128—classifyPromptingContextsetsoperatorInLoop;:369decideDeferenceStopHookActionconsumes it.isReportedMentionContext,isAttributiveCitationContext) are the precedent shape: narrow, per-phrase, falsifier-backed — not blanket.The Fix
Undecided by design — this ticket owns the decision, not a predetermined patch. Candidate shapes, each needing its own falsifier set:
operatorInLoop, on the ground that lane selection is never operator-owned. Risk: a genuine Tier-4 lane question ("should Clio take FM instead?") becomes a false positive.Option 2 is the incumbent and must be falsified before either other shape ships. The whole value of this hook is that it is a mirror, not a leash;
#14093records a carve-out proposed for this same hook that the detector already handled correctly, where implementing it "would have been net-negative."Decision Record impact
nonepending the decision; a shape that changes the block contract would want analigned-withnote against the §L3 no-hold substrate.Acceptance Criteria
#16325boundary spec instopHookDecision.spec.mjsis updated from pinning the current miss to pinning the new intended behavior.Out of Scope
#16325owns phrases).Avoided Traps
Measured false positives — 2026-08-11, the mirror of the case above
The Avoided Traps section of this ticket predicted this exact failure nine days ago — "a false positive against a real Tier-4 ask trains agents to route around the hook." It has now happened three times in one evening, so it is no longer a hypothetical cost to weigh against option 2.
All three are the MIRROR of the reproduction at the top of this ticket. That one is a false NEGATIVE in dialogue (
operatorInLoop=true→ blanket bypass → a real slip exempted). These are false POSITIVES on autonomous turns (operatorInLoop=false→ the matcher runs → it fires on the one thing §critical_gates #1 mandates). Both directions indict the same axis:operatorInLoopis not the discriminator, and it is wrong on both sides of itself.Instance 4 is the cleanest specimen available: the hook fired on a turn that was performing the escalation ladder correctly, in a message about the hook firing on correct escalation.
Why the existing carve cannot reach these.
isAttributiveCitationContextis anchored onper/as you said|directed|called— it exemptsyour callonly when it cites a prior operator decision (backward-looking). Every instance above is forward-looking: naming who owns a pending decision. There is no anchor to match, so the exemption is structurally unreachable for the mandated form.The docblock above
CITATION_ANCHORnames this class in full — "fired on the form §critical_gates #1 mandates — naming the gate that makes a merge the operator's decision" — while the code closes only the citation half of it. A docblock describing a class over an implementation covering one instance of it makes the gap invisible to the next reader.The cost is the training signal, not the noise (@neo-opus-grace's framing, which sharpens this ticket's own trap line): a hook that fires on correct Tier-4 routing teaches the seat to bury the handoff, phrase it evasively, or omit who owns the decision — producing exactly the silent-ownership failure the escalation ladder exists to prevent. And it would be invisible, because the hook goes quiet as the behaviour degrades.
What this does to the candidate shapes above: it does not decide between them, but it removes "the false-positive risk is speculative" as an argument for the incumbent. The reframing this ticket already proposes — whose decision is it, with merge/credentials/capacity/aesthetics staying carved — is the shape all four instances land inside, and those four domains are already enumerated in the escalation ladder rather than needing a new list.
Still unassigned and still Tier-4. These are L3 teeth gating every turn of every seat, so the disposition is the operator's; this section adds evidence, not a decision. Raised by @neo-opus-grace attacking the hook as the seat it gates, at her own request.
Related
#16325— the registry entry; its Out of Scope section defers this question here.#14093— prior carve-out proposal on this same hook, retracted after measurement.Live latest-open sweep: latest 20 open issues at 2026-08-02T10:40:17Z; no equivalent. A2A in-flight claim sweep: no
[lane-claim]on the deference detector or Stop-hook carve.Origin Session ID: eb230051-9e42-4e6b-b540-112a79accc3a
Retrieval Hint:
query_raw_memories("operatorInLoop carve deference blanket bypass lane selection decision ownership")— anchors:deferencePhraseMatch.mjs:130,stopHookDecision.mjs:128.Authored by Vega (Claude Opus 5, Claude Code) — unclaimed; filed to preserve the question, not to reserve the work.